SmolDocling vs Pixtral
Pick Pixtral if you need to answer questions over charts and documents, or already call the Mistral API. Pick SmolDocling only for experiments converting English pages on small hardware. Pixtral is a general 12B vision model with a published DocVQA score of 90.7, while SmolDocling is a 256M preview model that outputs DocTags for Docling. Both are no longer the maintainers' recommendation: Pixtral is marked deprecated and SmolDocling has been superseded by Granite-Docling.
SmolDoclinghuggingface/smoldocling | Pixtralmistral/pixtral | |
|---|---|---|
| Type | Open source | Vision model |
| Price | 1 credit / page | 2 credits / page |
| Per 1,000 pages | $1.00 | $2.00 |
| Typical speed | 1.0 s / page | 3.0 s / page |
| Best for | Very cheap page conversion | Questions about document images |
| Inputs | PDF, images | images |
| Languages | Mainly English | Multilingual |
| Max pages per request | 500 | 500 |
| Tables | ✓ Yes | No |
| Handwriting | No | No |
| Math | ✓ Yes | No |
| Bounding boxes | ✓ Yes | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | No | ✓ Yes |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose SmolDocling if
- It is 256M parameters, so it runs on modest hardware, and it takes about 0.35 seconds per page on an A100 with vLLM.
- It returns structured output with location tags: DocTags, Markdown, HTML, JSON, LaTeX and bounding boxes.
- One model handles text, tables, code, formulas and charts for English page conversion on a low budget.
Choose Pixtral if
- It scores 90.7 ANLS on DocVQA, so it suits question answering over documents and charts.
- You can call it through the Mistral API at a launch price of $0.15 per million input and output tokens, or self-host the Apache-2.0 weights.
- Its 128K token context allows many images in one request.
Key differences
| SmolDocling | Pixtral | |
|---|---|---|
| Model size | 256M parameters | 12.4B (12B decoder + 400M vision encoder) |
| How you run it | Self-host only | Mistral API (pixtral-12b-2409) or self-host open weights |
| Price | Free, self-hosted (you pay for compute) | $0.15 per million input and output tokens at API launch; free to self-host |
| Output formats | DocTags, Markdown, HTML, JSON, LaTeX, bounding boxes | Plain text, Markdown, JSON |
| Benchmarks | TEDS 0.82 on FinTabNet tables, full-page OCR F1 0.80, OCRBench 338 (all reported by the authors) | DocVQA 90.7 ANLS, ChartQA 81.8, MMMU 52.5 (reported by Mistral) |
| Languages | English only | Reported accuracy drop on non-Latin scripts, handwriting and small fonts |
| Known failure | Endless token repetition and missing closing tags; Docling issue #1515 reports output loops that make generation very slow | A launch thread user said it does not always get OCR anywhere near correct; Q4 quantization reportedly degrades small text OCR |
Sources
- SmolDocling model card
- Docling issue #1515
- Mistral announcement
- Mistral Sept 2024 pricing
- HN launch thread
What each one is
SmolDocling
Tiny 256M vision model from Hugging Face and IBM that outputs DocTags. Predecessor of Granite Docling.
SmolDocling detailsPixtral
Mistral's open-weight vision models, Pixtral 12B and Pixtral Large. Mistral's API no longer serves them, so they run self-hosted.
Pixtral detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "huggingface/smoldocling","fallbacks": ["mistral/pixtral"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
SmolDocling alternativesPixtral alternativesBest OCR models in 2026